Using interpretable machine learning to analyze the trajectory changes of serum albumin to predict the mortality rate of sepsis: a cohort study based on MIMIC-IV

Abstract

Background

Although hypoalbuminemia at ICU admission is associated with increased in-hospital mortality in septic patients, the prognostic value of serum albumin trajectory changes based on repeated measurements remains unclear. This study aimed to identify longitudinal serum albumin trajectory changes, evaluate their association with clinical outcomes at different temporal windows, and construct interpretable machine learning (ML) models for risk prediction.

Methods

Using the MIMIC-IV v3.0 database, trajectory analysis (via the traje R package) classified serum albumin trajectory changes into three groups: Persistently Low, Persistently Moderate, and Persistently High. Kaplan-Meier survival curves and Cox regression models were employed to analyze associations between trajectory groups and ICU, 28-day, and 90-day mortality. The Boruta algorithm screened feature variables, and seven ML models (XGBoost, SVM, LightGBM, RandomForest, e.g.) were developed. Model performance was evaluated using ROC curves, and SHapley Additive exPlanations (SHAP) values provided interpretability. A nomogram was constructed for clinical utility.

Results

Among 1,714 included patients, trajectory groups comprised Persistently Low (n = 477), Persistently Moderate (n = 749), and Persistently High (n = 488). Multivariable Cox regression (fully adjusted Model 3) revealed that compared to the Persistently Low group, both the Persistently High group and Persistently Moderate group exhibited significantly lower 28-day mortality risk (HR = 0.72, 95% CI: 0.56–0.93 for High; HR = 0.77, 95% CI: 0.61–0.96 for Moderate) and 90-day mortality risk (HR = 0.74, 95% CI: 0.60–0.92 for High; HR = 0.72, 95% CI: 0.60–0.87). For ML predictions, LightGBM achieved optimal AUCs for ICU mortality (0.758) and 28-day mortality (0.753), while RandomForest performed best for 90-day mortality (AUC = 0.721). SHAP analysis ranked serum albumin trajectories as the third most critical predictor for 90-day mortality. The nomogram showed good discrimination for 90-day mortality prediction, with an AUC of 0.765, and the calibration curve suggested acceptable agreement between predicted and observed risks.

Conclusion

Serum albumin trajectory changes in sepsis were associated with mortality outcomes at different time windows. In the fully adjusted Cox models, these trajectory changes remained significantly associated with 28-day and 90-day all-cause mortality, whereas their association with ICU mortality was attenuated after comprehensive adjustment. The integrated machine learning framework, particularly LightGBM and RandomForest, provides clinicians with interpretable tools for prognosis stratification in patients with sepsis.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12879-026-14034-7.

Introduction

Sepsis is a common syndrome in the Intensive Care Unit (ICU), primarily caused by an imbalance between anti-inflammatory and pro-inflammatory responses in the host following infection, thereby leading to multiple organ dysfunction [1], and it affects millions of people worldwide annually. Despite advancements in critical care research and the availability of sufficient ICU resources, sepsis remains a leading cause of death in ICUs worldwide, with persistently high mortality rates [2]. More critically, approximately 25% of survivors develop Persistent Inflammation, Immunosuppression, and Catabolism Syndrome (PICS), accompanied by sequelae such as cognitive impairment and sarcopenia, which elevate 1-year readmission and mortality risks, imposing substantial burdens on families and healthcare systems [3, 4]. In this context, establishing early-warning biomarker systems has become a critical priority for improving sepsis outcomes. Previous studies have confirmed that hypoalbuminemia is prevalent in septic patients, with serum albumin levels strongly correlating with prognosis—those with hypoalbuminemia exhibit significantly elevated mortality risks [5].

Serum albumin, a pleiotropic molecule with anti-inflammatory (via endotoxin binding), antioxidant (containing 17 sulfhydryl groups), and microcirculatory protective functions (by maintaining colloid osmotic pressure), plays a critical role in prognostic evaluation across diverse diseases. Recent studies increasingly recognize that dynamic albumin trajectories—rather than single time-point measurements—better predict survival and disease progression. For instance, a study in peritoneal dialysis patients revealed heterogeneous albumin trajectories (rising, stable, or declining during treatment), where improved albumin trajectories were inversely associated with mortality risk (HR = 0.68, 95% CI: 0.52–0.89), while baseline albumin levels showed no statistically significant association with outcomes [6]. These findings underscore the necessity of longitudinal albumin monitoring to guide clinical prognosis assessments. Thus, analyzing albumin trajectories in sepsis may enable early risk stratification and tailored therapeutic interventions to improve survival outcomes.

In recent years, machine learning (ML) has gained increasing traction in medical research, demonstrating exceptional performance in cancer diagnosis, survival prediction, and prognosis. Its application has also extended to mortality prediction in sepsis. ML models excel at processing large-scale clinical datasets, identifying latent risk factors, and delivering personalized prognostic insights. For example, a multicenter study leveraging ML algorithms analyzed electronic health records from sepsis patients across multiple hospitals, developing a prediction model that achieved an AUC of 0.90 for mortality prediction—significantly outperforming conventional scoring systems [7]. Furthermore, advancements in model interpretability, such as the SHAP (SHapley Additive exPlanations) framework, have enhanced clinical trust. A recent study employing SHAP analysis identified age, red blood cell distribution width (RDW), and respiratory rate as critical determinants of sepsis mortality. This interpretability not only bolsters clinical utility but also provides actionable insights for decision-making, bridging the gap between predictive analytics and bedside practice [8].

Methods and materials

Data

The data for this study were sourced from MIMIC-IV 3.0 (Medical Information Mart for Intensive Care), a publicly accessible, de-identified critical care database containing comprehensive clinical records, including demographics, laboratory parameters, comorbidities, medication histories, surgical procedures, disease severity scores (e.g., SOFA, APACHE II), and survival outcomes. Authorized data extraction was performed by a certified investigator (Certification ID: 44274909) following completion of the Collaborative Institutional Training Initiative program. As an anonymized public resource compliant with the Health Insurance Portability and Accountability Act, MIMIC replaces protected health information with randomized identifiers. Ethical approval for this study was granted by the Institutional Review Board of Binhai County People’s Hospital with a waiver of informed consent due to the retrospective and anonymized nature of the data.

Study population

The study cohort comprised adult patients admitted to the intensive care unit (ICU) with sepsis from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.0) database. Sepsis patients were identified strictly based on specific International Classification of Diseases, Ninth and Tenth Revisions (ICD-9 and ICD-10) diagnostic codes from the primary or secondary diagnosis fields. To catch the most prominent and high-frequency presentations of sepsis, we applied a customized code-matching list including ICD-9 codes (995.91, 995.92) and ICD-10 codes (A41.9, R65.20, R65.21, A41.01, A41.51, A41.59, A41.81, and A41.89). To eliminate confounding effects from repeated stays, only the patient’s first ICU admission was evaluated.

In this study design, patients were classified into the albumin trajectory group based on longitudinal albumin monitoring conducted on the 1st, 3rd, and 5th days after the diagnosis of sepsis. Only patients who survived until the 5th day and had sufficient albumin monitoring data to define their longitudinal change trajectory were included in the trajectory analysis. Therefore, patients who died before the 5th day were not included in the trajectory cohort.

Exclusion criteria (detailed in Supplementary Fig. S1) included: (1) age < 18 years; (2) missing critical laboratory or clinical data; and (3) death before the day-5 landmark or insufficient albumin data for trajectory classification. After applying these criteria, 1,714 critically ill sepsis patients were included in the final analysis.

Data extraction

Data extraction and management were performed using Structured Query Language (SQL) and PostgreSQL v15.0. We systematically retrieved the following variables: patient ID, age, sex, comorbidities (hypertension, diabetes mellitus, stroke, heart failure, renal injury), laboratory parameters (white blood cells, hemoglobin, platelets, triphasic albumin measurements, lactate, alanine aminotransferase [ALT], aspartate aminotransferase [AST], blood urea nitrogen, serum creatinine), disease severity scores (Sequential Organ Failure Assessment [SOFA], Acute Physiology Score III [APSIII]), infection site, ICU length of stay, mortality dates, and therapeutic interventions (albumin/glucocorticoid administration). Albumin levels were measured at three timepoints: day 1 (diagnosis confirmation), day 3, and day 5 post-sepsis diagnosis. All other laboratory biomarkers were captured within 24 h of sepsis diagnosis to ensure temporal alignment with acute pathophysiology.

Study outcome

The primary endpoint of this study is all-cause mortality, rather than specific mortality related to sepsis. The assessment time window is divided into three stages: (1) during hospitalization in the intensive care unit (ICU) (ICU mortality rate), (2) within 28 days after admission to the ICU (28-day mortality rate), and (3) within 90 days after admission to the ICU (90-day mortality rate). Defining these endpoint indicators strictly from the moment the patient is admitted to the ICU ensures a consistent baseline for conducting longitudinal survival analysis. Since the classification of albumin change trajectories requires albumin test results from the 5th day after the patient’s survival, all outcome analyses were conducted for patients who survived until the 5th day and had sufficient albumin data to determine the trajectory classification. Therefore, the results of ICU mortality, 28-day mortality, and 90-day mortality in the analysis reflect the mortality risk of the cohort meeting the 5-day survival trajectory criteria within these preset time windows, rather than the overall mortality risk of the sepsis patient population from the time of ICU admission.

Analysis of serum albumin trajectory changes

Trajectory analysis was conducted using the traj package in STATA v17.0, implementing the group-based trajectory modeling (GBTM) framework proposed by Leffondré et al. [9]. This method identified distinct serum albumin trajectory changes by clustering patients with similar longitudinal albumin profiles into subgroups. Because the three identified trajectories were relatively stable and approximately parallel, they should be interpreted as trajectory changes characterized mainly by persistent differences in albumin levels, rather than as sharply rising or falling trajectories. Model selection followed established criteria [10–12]: Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC), subgroup sample size adequacy (> 5% of cohort), and Odds of Correct Classification (OCC). The final trajectory model was selected through iterative optimization of these parameters.

Statistical analysis

The research group categorized patients based on the results of the albumin trajectory analysis into three groups: Persistently Low, Persistently Moderate, and Persistently High. Variables that did not conform to a normal distribution were expressed as median and interquartile range (IQR), and the Kruskal-Wallis test was used to compare differences among groups. Categorical variables were expressed as counts and percentages (%), and were analyzed using the chi-square test or Fisher’s exact test.

To evaluate the relationship between albumin trajectories and the three outcomes, Kaplan-Meier curves and log-rank tests were used to assess survival differences among albumin trajectory groups. Additionally, three multivariable Cox proportional hazards models were constructed to calculate adjusted hazard ratios (HRs) and 95% confidence intervals (CIs). Model 1 was unadjusted; Model 2 was adjusted for age, sex, SOFA score, and APSIII; and Model 3 was further adjusted for comorbidities, laboratory indicators, infection sites, and other clinically relevant variables based on Model 2. Subsequently, stratified analyses were performed to evaluate the association between albumin trajectory groups and outcomes across prespecified subgroups, including sex, hypertension, diabetes, heart failure, stroke, and renal injury-related variables.

Before machine learning model development, missingness of all candidate variables was systematically evaluated. Variables with a missingness rate exceeding 25% were excluded from the study. After applying the inclusion and exclusion criteria, the missingness rates of the retained candidate variables were relatively low. For variables with a missingness rate below 25%, missing values were imputed using a non-parametric random forest-based imputation method before dataset partitioning. The entire cohort was then divided into a training set and a validation set at a 7:3 ratio for machine learning model development and evaluation. We acknowledge that performing imputation before dataset partitioning may introduce a limited risk of information leakage by allowing the validation-set distribution to influence the preprocessing procedure. Given the low proportion of missing values among the retained variables after strict data screening, the practical impact of this issue was expected to be limited; nevertheless, this issue was considered when interpreting the machine learning results.

In this study, we developed seven machine learning (ML) models using binary classification to predict ICU all-cause mortality, 28-day all-cause mortality after ICU admission, and 90-day all-cause mortality after ICU admission in patients with sepsis. After missing data handling and dataset partitioning, feature selection was performed within the training set using the Boruta algorithm with 500 iterations. To reduce information redundancy and multicollinearity among the Boruta-selected variables, especially before translating them into an interpretable clinical regression framework, the variance inflation factor (VIF) was calculated within the training set, and features demonstrating a VIF ≥ 5 were sequentially excluded from subsequent modeling. The final selected feature set was then applied unchanged to the validation set.

We employed seven ML algorithms to construct predictive models: eXtreme Gradient Boosting (XGBoost), Gradient Boosting Decision Tree (GBDT), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost), Logistic Regression, Random Forest, and Support Vector Machine (SVM). To optimize generalizability and control for overfitting, model development and internal performance estimation were conducted using a 5-fold cross-validation strategy within the training set. For each algorithm, we used the default baseline hyperparameters implemented in the corresponding R packages as benchmark settings for clinical predictive modeling.

Model performance was comprehensively assessed on the validation set using the area under the receiver operating characteristic curve (AUC), along with sensitivity, specificity, recall, F1-score, and accuracy, to identify the optimal model. The SHapley Additive exPlanations (SHAP) method was applied to visualize feature contributions through bar plots and beeswarm plots, and to evaluate the impact of specific features on individual predictions, thereby enhancing the interpretability of model decisions. Finally, the top five features that contribute the most to the prediction of 90-day all-cause mortality, identified based on the SHAP values, were used to construct a nomogram. Since the SHAP analysis is only used for feature interpretation and ranking, the nomogram itself is constructed based on the multivariate logistic regression model.

To further assess whether limited missing-data handling influenced the machine learning interpretation, a complete-case sensitivity analysis was performed for the 90-day mortality prediction model by excluding patients with missing values in WBC, hemoglobin, platelets, creatinine, blood urea nitrogen, ALT, or AST. The 90-day mortality model was then rebuilt using the same machine learning framework, and SHAP-based feature importance was re-evaluated.

All statistical analyses were performed using R software (version 4.3.0) and STATA 17.0 (64 bits), and two-sided p-values < 0.05 were considered statistically significant.

Results

Construction of the serum albumin trajectory change model

The research team performed albumin trajectory analysis using the traje package in R software to identify trends of albumin levels across different time points. Following the parameter criteria including BIC (7533.48), AIC (7457.23), OCC, and group size distribution (Table 1), a three-group model was ultimately selected based on established selection criteria. Patients were classified into three distinct trajectories (Fig. 1): Group 1: Persistently Low – indicating sustained hypoalbuminemia throughout treatment in sepsis patients; Group 2: Persistently Moderate – maintaining intermediate albumin levels; Group 3: Persistently High – demonstrating consistently elevated albumin concentrations during treatment. As shown in Fig. 1, the three trajectory-change groups were mainly characterized by different albumin levels that remained relatively stable over time, rather than by clearly divergent increasing or decreasing slopes. Therefore, these groups represent serum albumin trajectory changes with persistent level separation, rather than strongly dynamic rising or falling trajectories.

Table 1.

Analysis results of albumin trajectory evaluation indicators

value	Group1	Group2	Group3
prob	0.278	0.426	0.296
CI lower	0.264	0.411	0.284
CI upper	0.293	0.441	0.307
group_rate	0.278	0.437	0.285
Ave PP	0.907	0.868	0.927
OCC	25.161	8.881	30.046

Fig. 1.

Group-based trajectory modeling (GBTM) of longitudinal serum albumin levels

Comparison of the basic characteristics and clinical information of the included patients

This study enrolled 1,714 sepsis patients meeting inclusion criteria, stratified into three albumin trajectory groups: Persistently Low (n = 477), Persistently Moderate (n = 749), and Persistently High (n = 488). Table 2 summarizes baseline characteristics and clinical parameters across trajectory groups. Comparative analyses revealed: Persistently Low group exhibited significantly higher white blood cells (WBC), platelets, blood urea nitrogen (BUN), and APSIII scores (all p < 0.05). Persistently High group showed the highest hemoglobin levels, while the Persistently Moderate group had elevated liver function markers (AST/ALT; both p < 0.05). Acute kidney injury prevalence was greatest in the Persistently Low group (p < 0.05), with no significant inter-group differences in other comorbidities. No statistically significant differences were observed in age, sex, creatinine (Cr), lactate (Lac), or SOFA scores, potentially attributed to the relatively limited sample size.

Table 2.

Incorporation of patients’ baseline characteristics

Characteristic	All	Persistently Low	Persistently Moderate	Persistently High	p value
N	1714	477	749	488
Age, years	67.00(56.00,79.00)	68.00(57.00, 79.00)	67.00(56.00, 79.00)	67.00(56.00, 78.00)	0.632
Gender, n (%)
Female	684 (39.9%)	195 (40.9%)	300 (40.1%)	189 (38.7%)	0.788
Male	1030 (60.1%)	282 (59.1%)	449 (59.9%)	299 (61.3%)
WBC 10^9/L	11.80(7.68,17.20)	12.80(7.90, 19.50)ab	11.30(7.50, 16.90)	11.25(7.60, 15.38)	0.001
Hb g/dL	10.10(8.40,11.80)	9.60(8.20, 11.40)ab	10.10(8.50, 11.70)	10.50(8.40, 12.38)	< 0.001
PLT 10^9/L	150(90,223.25)	167(92,2520)b	151(93,220)	138(83,205.75)	0.001
CR mg/dL	1.30(0.80,2.10)	1.30(0.80, 2.00)	1.20(0.80, 2.00)	1.30(0.90, 2.30)	0.124
BUN mg/dL	26.00(15.00,45.00)	29.00(17.00,49.00)a	25.00(15.00, 40.00)	25.00(15.00, 46.00)	0.006
ALT IU/L	40.00(20.00,135.00)	35.00(18.00,99.00)a	47.00(20.00,203.00)	36.00(20.0,124.75)	0.001
AST IU/L	70.00(33.00,227.00)	61.00(29.50,157.00)ab	74.00(34.00,285.50)	68.50(38.0,226.00)	0.004
Lac mmol/L	2.30(1.50,3.40)	2.20(1.50, 3.30)	2.40(1.50, 3.60)	2.30(1.60, 3.10)	0.379
SOFA	8.00(5.00,11.00)	8.00(5.00, 11.00)	8.00(5.00, 10.50)	8.00(5.00, 11.00)	0.943
APSIII	58.00(44.00,74.00)	63.00(51.00, 79.00)ab	56.00(43.00, 72.00)	54.00(39.00, 74.75)	< 0.001
Albumin g	12.50(0.00,128.12)	25.00(0.00, 125.00)	12.50(0.00, 112.50)	25.0(0.00, 196.88)a	0.031
Hypertension, n (%)					0.876
No	1161 (67.7%)	327 (68.6%)	507 (67.7%)	327 (67.0%)
Yes	553 (32.3%)	150 (31.4%)	242 (32.3%)	161 (33.0%)
Diabetes, n (%)					0.251
No	1248 (72.8%)	361 (75.7%)	536 (71.6%)	351 (71.9%)
Yes	466 (27.2%)	116 (24.3%)	213 (28.4%)	137 (28.1%)
Heart failure, n (%)					0.562
No	1283 (74.9%)	365 (76.5%)	559 (74.6%)	359 (73.6%)
Yes	431 (25.1%)	112 (23.5%)	190 (25.4%)	129 (26.4%)
Kidney Injury, n (%)					< 0.001
No	539 (31.4%)	116 (24.3%)	255 (34.0%)	168 (34.4%)
Yes	1175 (68.6%)	361 (75.7%)	494 (66.0%)	320 (65.6%)
Stroke, n (%)					0.287
No	1608 (93.8%)	452 (94.8%)	705 (94.1%)	451 (92.4%)
Yes	106 ( 6.2%)	25 ( 5.2%)	44 ( 5.9%)	37 ( 7.6%)
Respiration, n (%)					0.247
No	1332 (77.7%)	368 (77.1%)	572 (76.4%)	392 (80.3%)
Yes	382 (22.3%)	109 (22.9%)	177 (23.6%)	96 (19.7%)
Coagulation, n (%)					0.002
No	915 (53.4%)	283 (59.3%)	397 (53.0%)	235 (48.2%)
Yes	799 (46.6%)	194 (40.7%)	352 (47.0%)	253 (51.8%)
Liver, n (%)					0.005
No	994 (58.0%)	303 (63.5%)	431 (57.5%)	260 (53.3%)
Yes	720 (42.0%)	174 (36.5%)	318 (42.5%)	228 (46.7%)
Cardiovascular, n (%)					< 0.001
No	733 (42.8%)	171 (35.8%)	311 (41.5%)	251 (51.4%)
Yes	981 (57.2%)	306 (64.2%)	438 (58.5%)	237 (48.6%)
CNS, n (%)					0.031
No	1298 (75.7%)	376 (78.8%)	572 (76.4%)	350 (71.7%)
Yes	416 (24.3%)	101 (21.2%)	177 (23.6%)	138 (28.3%)
Renal, n (%)					0.486
No	879 (51.3%)	254 (53.2%)	384 (51.3%)	241 (49.4%)
Yes	835 (48.7%)	223 (46.8%)	365 (48.7%)	247 (50.6%)
Glucocorticoids, n (%)					0.767
No	1058 (61.7%)	301 (63.1%)	458 (61.1%)	299 (61.3%)
Yes	656 (38.3%)	176 (36.9%)	291 (38.9%)	189 (38.7%)
Albumin therapy, n (%)					0.316
No	826 (48.2%)	217 (45.5%)	374 (49.9%)	235 (48.2%)
Yes	888 (51.8%)	260 (54.5%)	375 (50.1%)	253 (51.8%)

Note: Continuous variables are expressed as median (interquartile range, IQR) and compared using the Kruskal-Wallis test; categorical variables are expressed as counts with percentages (%) and analyzed using the chi-square test or Fisher’s exact test. Post-hoc pairwise multiple comparisons were conducted using Dunn’s test with Bonferroni correction for continuous variables demonstrating an overall p < 0.05

a: p < 0.05$ versus the Persistently Moderate group

b: p < 0.05$ versus the Persistently High group

ab: p < 0.05$ versus both the Persistently Moderate and Persistently High groups

SOFA: sequential organ failure assessment; APSIII: Acute Physiology and Chronic Health Evaluation III; WBC: White Blood Cell; Hb: Hemoglobin; PLT: Platelet; CR: Creatinine; BUN: Blood Urea Nitrogen; Lac: Lactic acid ; ALT: Alanine Aminotransferase; AST: Aspartate Aminotransferase

The relationship between albumin trajectory and mortality within the ICU in patients with sepsis

Based on albumin trajectory groups, Kaplan-Meier survival curves were generated to compare ICU mortality risk during hospitalization (Fig. 2A). The analysis revealed a trend toward the lowest survival rate in the Persistently Low group and the highest in the Persistently High group, though without statistical significance (log-rank P = 0.059). This trend aligned with mortality counts across groups, where the Persistently Low group had the highest number of deaths (Fig. 2D). Three multivariate Cox regression models (Table 3) were constructed to assess the stability of associations between albumin trajectories and ICU mortality: Model 1 (Unadjusted): Compared to the Persistently Low group, the Persistently Moderate group showed a non-significant reduction in mortality risk (HR = 0.84, 95% CI 0.63–1.13, P = 0.248), while the Persistently High group had a significantly lower risk (HR = 0.67, 95% CI 0.48–0.93, P = 0.018). Model 2 (adjusted for age, gender and severity score): It was found that the significant protective effect observed in the persistent high-risk group remained statistically robust, indicating an independent reduction in the risk of death (HR = 0.66, 95% CI 0.47–0.92, P = 0.014). Model 3 (Further adjusted for comorbidities and lab parameters): Neither group demonstrated statistical significance, though directional consistency with prior models was maintained. Subgroup analyses (Supplementary Table S1) confirmed lower mortality risks in the Persistently Moderate and High groups compared to Persistently Low, reinforcing the robustness of the observed trends.

Fig. 2.

Survival analysis and clinical mortality profiles across distinct albumin trajectory groups. (A–C) Kaplan-Meier curves with log-rank tests evaluating all-cause mortality across three distinct temporal windows: ICU stay (A), 28 days after ICU admission (B), and 90 days after ICU admission (C); (D) Statistical breakdown of day-specific raw mortality events, percentages, and baseline indicators across the three cohorts

Table 3.

Cox proportional hazards model analysis of albumin trajectory grouping and mortality outcomes in sepsis patients

Outcome	Model 1			Model 2			Model 3
	HR (95%CI)	P	P for trend	HR (95%CI)	P	P for trend	HR (95%CI)	P	P For trend
ICU all-cause mortality after ICU admission
Persistently Low	Reference		< 0.001	Reference		< 0.001	Reference		< 0.001
Persistently Moderate	0.84 (0.63–1.13)	0.248		0.84 (0.63–1.12)	0.223		0.95 (0.70–1.30)	0.758
Persistently High	0.67 (0.48–0.93)	0.018		0.66 (0.47–0.92)	0.014		0.74 (0.52–1.06)	0.103
28-day all-cause mortality after ICU admission
Persistently Low	Reference		< 0.001	Reference		< 0.001	Reference		< 0.001
Persistently Moderate	0.64 (0.51–0.79)	< 0.001		0.64 (0.51–0.79)	< 0.001		0.77 (0.61–0.96)	0.021
Persistently High	0.60 (0.47–0.77)	< 0.001		0.60 (0.47–0.77)	< 0.001		0.72 (0.56–0.93)	0.013
90-day all-cause mortality after ICU admission
Persistently Low	Reference		< 0.001	Reference		< 0.001	Reference		< 0.001
Persistently Moderate	0.61 (0.51–0.74)	< 0.001		0.61 (0.51–0.74)	< 0.001		0.72 (0.60–0.87)	0.001
Persistently High	0.63 (0.51–0.77)	< 0.001		0.63 (0.51–0.77)	< 0.001		0.74 (0.60–0.92)	0.006

Model 1 was unadjusted

Model 2 was adjusted for age, sex and severity scores (sequential organ failure assessment, Acute Physiology and Chronic Health Evaluation III)

Model 3 was Model 2 further adjusted for comorbidities (heart failure, hypertension, diabetes, chronic kidney disease, acute renal failure, and stroke), laboratory results (WBC, Hb, PLT, AST, ALT, CR, BUN and Lac etc.), Infection site (respiration, coagulation, liver, cardiovascular, CNS and renal)

CI, confidence interval; HR, hazard ratio; ICU, intensive care unit

Because HRs and 95% CIs were rounded to two decimal places, some estimates in Model 1 and Model 2 appeared identical after rounding. We rechecked and reran the Cox models, and the covariates in Model 2 were correctly included. The similarity between Model 1 and Model 2 suggests that adjustment for age, sex, SOFA score, and APSIII had minimal influence on the association between albumin trajectory groups and mortality outcomes

Association of albumin trajectory and mortality at 28 days after ICU in patients with sepsis

For the analysis of 28-day all-cause mortality post-ICU admission, Kaplan-Meier survival curves were generated to compare mortality across albumin trajectory groups (Fig. 2B). The analysis demonstrated statistically significant differences (log-rank P < 0.001), with the Persistently Low group exhibiting the highest mortality, the Persistently High group the lowest, and the Persistently Moderate group intermediate. This aligned with mortality counts across groups (Fig. 2D). Multivariate Cox regression models (Table 3) further assessed the association: Model 1 (Unadjusted): Compared to Persistently Low, the Persistently High group had the lowest 28-day mortality risk (HR = 0.60, 95% CI 0.47–0.77, P < 0.001), while the Persistently Moderate group showed intermediate risk (HR = 0.64, 95% CI 0.51–0.79, P < 0.001). Model 2 (Adjusted for age, sex, severity scores): Persistently High maintained significantly lower risk (HR = 0.60, 95% CI 0.47–0.77, P < 0.001), with Persistently Moderate remaining intermediate (HR = 0.64, 95% CI 0.51–0.79, P < 0.001). Model 3 (Additional adjustments for comorbidities/lab parameters): Risk reduction persisted for Persistently High (HR = 0.72, 95% CI 0.56–0.93, P = 0.013) and Persistently Moderate (HR = 0.77, 95% CI 0.61–0.96, P = 0.021), though attenuated. Subgroup analyses (Supplementary Table S1) revealed significant interactions between albumin trajectories and specific comorbidities (e.g., hypertension, diabetes, heart failure), with the Persistently Low group consistently demonstrating the highest mortality risk across strata (P < 0.05).

Association of albumin trajectory and mortality at 90 days after ICU in patients with sepsis

For the analysis of 90-day all-cause mortality post-ICU admission, Kaplan-Meier survival curves (Fig. 2C) demonstrated statistically significant differences across albumin trajectory groups (log-rank P < 0.001), with the Persistently Low group showing the lowest survival, Persistently High the highest, and Persistently Moderate intermediate. Mortality counts aligned with these trends (Fig. 2D). Three multivariate Cox regression models (Table 3) were constructed to evaluate the association: Model 1 (Unadjusted): Compared to Persistently Low, the Persistently High group had the lowest 90-day mortality risk (HR = 0.63, 95% CI 0.51–0.77, P < 0.001), followed by Persistently Moderate (HR = 0.61, 95% CI 0.51–0.74, P < 0.001). Model 2 (Adjusted for demographics/severity): Persistently High maintained the lowest risk (HR = 0.63, 95% CI 0.51–0.77, P < 0.001). Model 3 (comprehensive adjustment model): It was found that the 90-day mortality risk was significantly reduced in the group with consistently high albumin levels (risk ratio = 0.74, 95% confidence interval 0.60–0.92, P = 0.006) and in the group with consistently moderate albumin levels (risk ratio = 0.72, 95% confidence interval 0.60–0.87, P = 0.001). Although the effect size slightly weakened in the direction of zero after controlling for these extensive clinical covariates, the dynamic albumin trajectory group still has a highly reliable and independent predictive value for the long-term survival of patients with sepsis. Subgroup analyses (Supplementary Table S1) confirmed significant interactions between albumin trajectories and comorbidities (e.g., hypertension, diabetes, heart failure), with mortality risk gradients (Persistently Low > Moderate > High) remaining statistically robust (P < 0.05).

Characteristic preselection of the ML model

For the all-cause mortality rate of patients with septicemia occurring in the ICU, we analyzed the all-cause mortality rate within 28 days and 90 days after ICU admission. This study included a total of 27 variables. To filter out variables with a high contribution weight, we used the Boruta algorithm to select relevant features (Fig. 3). First, we analyzed the factors associated with mortality during ICU admission (Fig. 3A), which included variables such as APSIII, SOFA, AST, ALT, BUN, Lac, CR, Albtotalval, and PLT. For the analysis of mortality occurring within 28 days after ICU admission (Fig. 3B), the variables included APSIII, SOFA, AST, ALT, BUN, Lac, CR, Albtotalval, PLT, and group (albumin trajectory). For the analysis of mortality occurring within 90 days after ICU admission (Fig. 3C), the variables included APSIII, SOFA, AST, ALT, BUN, Lac, CR, Albtotalval, PLT, Hb, renal injury, urinary tract infection, and group (albumin trajectory). We found that the albumin trajectory factor was not included in the outcome of mortality during ICU admission, while it was included in the other two outcomes.

Fig. 3.

Feature importance and variable screening leveraging the Boruta algorithm. Multi-variable feature selection distributions executed for the three clinical endpoints: ICU all-cause mortality (A), 28-day all-cause mortality after ICU admission (B), and 90-day all-cause mortality after ICU admission (C)

ML model construction and evaluation

Based on the characteristics selected above, we constructed seven machine learning models to identify the risk factors for three outcomes in patients with sepsis (Fig. 4). For the analysis of mortality during ICU admission, the ROC curve (Fig. 4A) and other metrics (Fig. 4D) were evaluated across all ML models, with Light GBM being considered the best model, having the highest AUC (0.758) and F1 score (0.747), outperforming other models. For the analysis of all-cause mortality 28 days after ICU admission, the ROC curve (Fig. 4B) and other metrics (Fig. 4E) were evaluated across all ML models, where Light GBM remained the best model, with the highest AUC (0.753) and F1 score (0.741). For the analysis of all-cause mortality 90 days after ICU admission, the ROC curve (Fig. 4C) and other metrics (Fig. 4F) were evaluated across all ML models, with Random Forest being considered the best model, having an AUC (0.721) and F1 score (0.704).

Fig. 4.

Predictive performance matrix of seven machine learning models across three endpoints. (A–C) Receiver operating characteristic (ROC) curves evaluated on the independent validation split for ICU mortality (A), 28-day mortality (B), and 90-day mortality (C); (D–F) Companion evaluation metrics (comprising AUC, Accuracy, Sensitivity, Specificity, F1-score, and Recall) for ICU mortality (D), 28-day mortality (E), and 90-day mortality (F)

Visualization of feature importance

Figure 5 shows a comprehensive population map and a hive plot of three outcomes for patients. First, an analysis of mortality during ICU admission was conducted, where we ranked the corresponding SHAP values of each feature in the Light GBM model validation cohort (Fig. 5A, B). The top variables mainly included BUN, APSIII, and PLT. Additionally, Fig. 5G provides a detailed case study of a patient, illustrating the model’s prediction process for a specific patient. In this visualization, red indicators represent negative contributions to the prediction, while blue indicates positive impacts. The f(x) value represents the actual SHAP value of each factor. Further analysis of all-cause mortality 28 days after ICU admission calculated and ranked the corresponding SHAP values of features in the Light GBM model validation cohort (Fig. 5C, D). The albumin trajectory ranked lower, contributing less. Figure 5H presents a case study of a patient. Finally, an analysis of all-cause mortality 90 days after ICU admission calculated and ranked the corresponding SHAP values of each feature in the Random Forest model validation cohort (Fig. 5E, F). The albumin trajectory ranked third in importance. To better guide clinical decision-making and demonstrate the practical application value of our predictive framework, we used SHAP importance plots to conduct a detailed analysis of a specific case in the 90-day mortality model (Fig. 5I). In this visual chart, the baseline values represent the average predicted probability of 90-day mortality for this group, while the red and blue indicators show the characteristics that increase or decrease the individual risk. This particular sepsis patient exhibited multiple high-risk acute symptoms upon admission to the intensive care unit, including elevated blood urea nitrogen levels (BUN = 31.0 milligrams per deciliter), high serum lactate levels (Lac = 5.2 millimoles per liter) indicating acute tissue hypoperfusion, and severe underlying hypoalbuminemia (Albtotalval = 20.0 g per liter). Driven by these significant red arrows, the initial mortality risk of this patient was severely pushed towards an unfavorable prognosis direction. However, compensatory blue arrows including a low Acute Physiology Score III (APSIII = 27.0) and mild organ dysfunction (SOFA = 1.0) provided protective buffering effects. Crucially, the ultimate determining factor for the patient’s survival lies in their classification into the 3.0 group (the persistent high albumin trajectory group). Notably, classification into the Persistently High albumin trajectory group contributed to a lower predicted 90-day mortality risk in this individual case. However, this factor should be interpreted as one component of the model rather than the ultimate determinant of survival. Its protective contribution appeared to partially offset, but not completely override, the effects of acute metabolic stressors and severity-related indicators.

Fig. 5.

Machine learning model interpretability via SHapley Additive exPlanations (SHAP) visualization. (A, B, G) SHAP beeswarm plot, summary feature bar plot, and individual force plot prediction (f(x) path) for the ICU mortality endpoint; (C, D, H) Summary evaluations and individual decision visualization for the 28-day mortality endpoint; (E, F, I) Feature hierarchies and individual clinical patient simulation for the 90-day mortality endpoint, highlighting the trajectory group as a top-3 long-term predictor

To further evaluate whether limited missing laboratory data influenced the model interpretation, we performed a complete-case sensitivity analysis by excluding patients with missing values in WBC, Hb, PLT, creatinine, BUN, ALT, or AST, and then rebuilt the 90-day mortality prediction model using the same machine learning framework (Supplementary Fig. S2). The complete-case model achieved an AUC of 0.699. In the SHAP-based feature importance analysis, APSIII, lactate, SOFA, albumin trajectory group, and hemoglobin remained the five leading contributors. Notably, the albumin trajectory group remained among the major predictive variables, ranking fourth in mean absolute SHAP value. These findings suggest that the contribution of albumin trajectory changes to 90-day mortality prediction was not substantially driven by limited missing-data handling.

Construction of the nomograms

To enhance the clinical applicability of the model and facilitate rapid decision-making by clinicians, we selected the top five variables for plotting a nomogram (Fig. 6) based on the best model constructed above and the variable importance from SHAP analysis. Our previous research found that the trajectory of albumin significantly contributes to the all-cause mortality rate 90 days after ICU admission, while its contribution to the other two outcomes is relatively low. Therefore, we only plotted the nomogram predicting the all-cause mortality rate 90 days after ICU admission. This nomogram was constructed based on a multivariable logistic regression model rather than directly derived from a machine learning model. The final logistic regression model included APSIII, hemoglobin, lactate, SOFA score, and albumin trajectory group. The albumin trajectory group was treated as a categorical variable, with the Persistently Low trajectory group set as the reference category. The coefficients, odds ratios, 95% confidence intervals, and P values of the final logistic regression model are provided in Supplementary Table S2. Figure 6A presents the nomogram of the predictive model. Figure 6B shows the calibration curve, which visually indicates acceptable agreement between the predicted probabilities and the observed 90-day mortality risk. Figure 6C and D display the ROC curves for each predictive factor and the overall model, respectively. The ROC analysis showed that the nomogram had good discriminative performance, with an AUC of 0.765.

Fig. 6.

Development, calibration, and discrimination assessment of the 90-day all-cause mortality prediction nomogram. (A) Nomogram constructed using the top five predictors identified from the optimal model and SHAP analysis. (B) Calibration curve showing the agreement between predicted and observed 90-day mortality risk. (C) ROC curves for individual predictors. (D) ROC curve of the overall nomogram model, demonstrating discrimination performance with an AUC of 0.765

Discussion

This study focuses on the changes in albumin trajectories of patients with sepsis through a retrospective cohort analysis combined with machine learning techniques. The main finding is that albumin trajectories are significantly associated with the prognosis of sepsis patients. The analysis shows that, using the Persistently Low group as a reference, the Persistently High group can significantly reduce the risk of death, while the Persistently Moderate group has a death risk between the two groups, and albumin trajectories are an independent factor for patient prognosis. Subsequently, seven ML algorithms were used to construct predictive models for the three prognoses of patients. The results indicate that the Light GBM and Random Forest models exhibit strong performance with good evaluation metrics. To gain a deeper understanding of the model, we used the SHAP method to visualize the three prognoses separately. The study found that albumin trajectories have a significant contribution to the 90-day all-cause mortality rate after ICU admission, while the contributions to the other two prognoses are insufficient. To provide clinicians with a more intuitive prediction of the 90-day all-cause mortality rate after ICU admission, we selected the top five variables to create a nomogram.

The three albumin trajectories identified in this study were relatively stable within the first 5 days after the diagnosis of sepsis and were roughly parallel. Therefore, the “trajectory changes” in this study mainly refer to the longitudinal dynamic characteristics of albumin levels that show a continuous separation, rather than significant upward or downward trends. This distinction is important because the prognostic information captured by the trajectory model may reflect the cumulative burden of persistent hypoalbuminemia or stable albumin status during the early stage of sepsis, rather than just rapid fluctuations over time.

The findings of this study significantly echo and complement previous literature. Existing research has shown that hypoalbuminemia is prevalent in critically ill patients, particularly among those with sepsis, and is negatively correlated with disease severity [13]. Hypoalbuminemia not only serves as a marker of inflammatory response and capillary leakage in septic patients but is also closely associated with coagulation disorders, immunosuppression, and multi-organ failure [14–16]. This study used trajectory modeling to demonstrate that serum albumin trajectory changes based on repeated measurements provide prognostic information beyond single-timepoint albumin assessment. This aligns with the conclusions of Kendall et al. [16] in their 2019 study, which identified albumin trend changes, admission albumin levels, and the lowest albumin levels as unique predictors of mortality. In the fully adjusted Cox model, the changes in serum albumin trajectories were significantly associated with 28-day and 90-day all-cause mortality, while the association with ICU mortality was weakened and did not reach statistical significance. However, since these estimates were derived from different outcome-specific models, the differences in p-values should not be interpreted as formal evidence that the albumin trajectory predicts a particular death endpoint more accurately than another. This discrepancy may be attributed to the fact that dynamic albumin changes reflect patients’ long-term nutritional status, chronic inflammatory persistence, and organ functional reserve [17, 18]. Patients in the persistently low albumin group exhibited significantly higher urea nitrogen levels (p < 0.05) and a greater proportion of renal injury (p < 0.05), suggesting that this trajectory may serve as a biomarker for chronic organ dysfunction.

The impact of dynamic albumin trajectories on prognosis may be mediated through the following interconnected pathways: First, persistently low albumin levels may reflect sustained catabolic imbalance in patients under inflammatory states. The essence of sepsis lies in its “inflammatory storm,” where systemic inflammatory responses exacerbate muscle protein breakdown. Concurrently, inflammatory cytokines (e.g., IL-6, TNF-α) significantly suppress albumin synthesis by inhibiting the hepatocyte nuclear factor-1α (HNF-1α) signaling pathway [19]. This persistent suppression of synthesis may explain why the Persistently Low group showed significantly increased 90-day mortality rather than short-term prognosis. Second, beyond metabolic regulation, albumin’s molecular properties directly influence oxidative stress balance. Its antioxidant and immunomodulatory functions likely play critical roles. Studies demonstrate that sulfhydryl groups on albumin molecules effectively scavenge oxygen free radicals, with albumin concentration positively correlating with total serum antioxidant capacity [20]. Thus, in the Persistently Low trajectory group, reduced antioxidant capacity may exacerbate endothelial oxidative damage, promoting microcirculatory dysfunction and interstitial organ edema. Additionally, albumin exerts vital immunomodulatory effects by binding circulating endotoxins and stabilizing protective immune signaling cascades; consequently, a persistently low albumin trajectory serves not merely as a passive biomarker of severity, but acts as a pathological culprit that facilitates a state of “immune paralysis,” inherently elevating the risks of secondary healthcare-associated infections [21]. Finally, the higher urea nitrogen levels and renal injury rates observed in our Persistently Low cohort suggest that albumin dynamics interact closely with renal hemodynamic mechanisms. Hypoalbuminemia-driven declines in systemic colloid osmotic pressure alter renal perfusion and activate the renin-angiotensin-aldosterone system, while ongoing longitudinal albumin depletion serves as an independent culprit that accelerates acute and subacute renal injury pathways [22]. These mechanisms imply that albumin trajectories do not merely echo systemic severity but actively participate in the multi-organ injury cascade, offering rational clinical intervention windows.

For the construction of prognosis models in septic patients, our research group employed multiple machine learning algorithms for modeling. This study innovatively incorporated albumin trajectories into the predictive feature system. Although albumin trajectories were not selected as key variables for predicting in-ICU mortality or 28-day post-ICU mortality, they significantly contributed to 90-day post-ICU mortality prediction (ranking third in SHAP values). This indicates that single-timepoint indicators, such as disease severity scores, are better at capturing acute-phase risks, while dynamic trajectory metrics hold unique value in predicting long-term disease evolution. This aligns with the “spatiotemporal heterogeneity of longitudinal biomarkers” theory proposed by researchers [10]. The superior performance of Light GBM in short-term prediction may stem from its efficient gradient-boosting mechanism in capturing complex interactions among laboratory indicators, whereas Random Forest’s advantage in medium-to-long-term prediction likely relates to its ensemble learning capability for time-dependent variables. Subsequently, to visualize the predictive model, we developed a nomogram for 90-day all-cause post-ICU mortality. Based on the optimal model and SHAP analysis of variable importance, the top five predictors were selected: APSIII, Hb, albumin trajectory (group), Lac, and SOFA. Consistent with prior studies [16, 23, 24], APSIII, Lac, and SOFA showed positive correlations with 90-day mortality risk (higher values associated with increased risk), while Hb and albumin trajectory exhibited inverse relationships (higher values linked to reduced risk). Although indicators such as WBC, PLT, comorbidities (hypertension, diabetes, renal injury), and infection sites were not included in the model, they remain clinically relevant. For instance, infections in the respiratory, digestive, or urinary systems significantly elevate 90-day post-ICU all-cause mortality [25, 26]. Additionally, concomitant heart failure markedly increases mortality risk. Studies have shown that septic patients developing heart failure during treatment face substantially higher mortality risks [27]. These risk factors warrant clinical attention, as early interventions targeting them may improve survival outcomes.

To further contextualize our findings, it is essential to compare our predictive framework with existing cohort studies that leveraged ML to forecast ICU mortality in sepsis patients. In recent years, numerous models utilizing large-scale retrospective electronic health records (such as the MIMIC or eICU databases) have achieved outstanding mathematical metrics by deploying advanced boosting trees or deep learning architectures. For instance, several benchmark models rely strictly on high-dimensional, cross-sectional physiological snapshots captured within the first 24 h of ICU admission to predict short-term mortality. While these frameworks exhibit excellent performance during hyper-acute resuscitation phases, they are inherently constrained by temporal blindness failing to register the host’s longitudinal adaptive responses or metabolic trajectories over time. Our study effectively addresses this gap by intertwining group-based trajectory modeling (GBTM) with ML frameworks. By demonstrating that a multi-day dynamic indicator (the serum albumin trajectory) possesses superior predictive capacities for long-term (90-day) survival compared to immediate, cross-sectional baseline parameters, our model introduces a shift from single-timepoint risk capturing to longitudinal trend tracking.

Recent studies and reviews have presented new perspectives, suggesting that sepsis should not be regarded as a unified clinical syndrome but rather as a heterogeneous disease that requires dynamic phenotypic analysis, precise risk stratification, and decision support tools with clinical interpretability [28]. It is particularly noteworthy that recent studies have emphasized the potential value of subphenotype analysis and diagnostic strategies for sepsis in identifying clinically significant patient subgroups; at the same time, it is pointed out that if the methods based on artificial intelligence are transparent, have been externally verified and can be clinically applied, they can effectively enhance the ability of early diagnosis, the level of risk stratification, and the formulation of individualized treatment plans [29]. At the same time, multi-center research evidence indicates that some previously proposed sepsis phenotypes may not generalize well in different independent cohorts, highlighting the importance of seeking a balance between algorithm complexity, clinical robustness, and external validity [30]. In this context, our framework was designed to integrate dynamic albumin trajectories into a clinically interpretable prognostic model rather than relying solely on high-dimensional “black-box” prediction. We first used the Boruta algorithm to select relevant features and then applied SHAP analysis to quantify and visualize the contribution of individual variables. Because conventional linear models may fail to capture the prognostic information embedded in dynamic longitudinal biomarkers, the random forest architecture enabled us to model nonlinear interactions involving albumin trajectory groups. Notably, albumin trajectory emerged as the third most important predictor of 90-day mortality, suggesting that longitudinal albumin patterns provide incremental prognostic information beyond static baseline measurements. Importantly, by translating these algorithmic findings into a concise five-variable clinical nomogram, our study provides an intuitive and interpretable tool for long-term prognostic stratification in patients with sepsis. However, as shown in the individual SHAP decision plot (Fig. 5I), even among patients with a favorable albumin trajectory, the absolute baseline risk was still largely driven by established severity-related indicators such as APS III and lactate. Therefore, although our framework successfully incorporates a dynamic trajectory-based biomarker, its clinical role should be regarded as complementary to, rather than a replacement for, traditional acute severity scores. The model should therefore be interpreted with appropriate clinical caution and further validated in independent multicenter cohorts.

The clinical implications of this study are reflected in three key aspects: First, dynamic albumin monitoring should be incorporated into the management pathways for septic patients and must not be overlooked. Critical time windows for monitoring—particularly on post-diagnosis days 1, 3, and 5—hold significant prognostic relevance. Second, visualization tools provided by machine learning models (e.g., SHAP honeycomb plots) enable clinicians to intuitively identify high-risk patient feature combinations. Finally, the developed nomogram offers a practical tool for rapid bedside risk assessment, with albumin trajectories and APSIII serving as core predictive elements.

Limitations

In this study, although a prognostic model for patients with sepsis was constructed using dynamic trajectory analysis and machine learning methods, several limitations should be acknowledged. First, due to the retrospective nature of the study, the influence of treatment-related interventions, such as exogenous albumin infusion, on trajectory classification could not be fully controlled. Second, the MIMIC-IV database lacks detailed information on nutritional support, which may have resulted in the omission of important confounding factors affecting albumin metabolism. Third, this study mainly focused on all-cause mortality at 28 and 90 days, while longer-term functional outcomes after critical illness were not evaluated. Fourth, we did not further describe or analyze the dose, timing, and frequency of albumin infusion, which may also be important factors influencing patient prognosis. Finally, the machine learning model was developed and internally validated using data mainly derived from a single institution in the MIMIC-IV database, and external validation in geographically independent cohorts was lacking. In addition, although variables with high missingness were excluded and the remaining variables had relatively low missingness after strict cohort screening, missing-value imputation was performed before splitting the cohort into training and validation sets. This preprocessing strategy may have introduced limited information leakage by allowing the validation-set distribution to influence the imputation procedure. Moreover, if feature selection procedures, such as Boruta screening and VIF filtering, are not fully confined to the training set, model performance may be optimistically estimated. Although two safeguards were applied to reduce the risk of overfitting, including a 5-fold cross-validation framework during model training and an unchanged 30% validation set for internal validation, systematic bias and potential information leakage cannot be completely excluded. Therefore, the machine learning results should be interpreted cautiously, and the generalizability and robustness of the model require further verification in geographically independent and multicenter external cohorts.

Future studies should focus on the following directions. First, multicenter prospective cohort studies or large independent external datasets, such as the eICU Collaborative Research Database or regional multicenter electronic medical records, should be used to externally validate the reliability and generalizability of trajectory-based predictive models. Second, metabolomics approaches could be integrated to further explore the molecular characteristics of different albumin trajectory subgroups. Third, real-time risk warning systems based on dynamic biomarkers should be developed and incorporated into electronic medical record systems to support individualized risk stratification and clinical decision-making.

Conclusion

This study integrates longitudinal trajectory analysis with machine learning modeling to evaluate the association between serum albumin trajectory changes and sepsis prognosis. The findings suggest that Persistently Moderate and Persistently High albumin trajectory groups were associated with lower 28-day and 90-day all-cause mortality compared with the Persistently Low group, while the association with ICU mortality was attenuated after full adjustment. Furthermore, a predictive framework combining repeated-measurement-derived biomarker information and clinical characteristics was established. These findings provide practical support for risk stratification in patients with sepsis, but they should not be interpreted as formal evidence that albumin trajectories predict 90-day mortality better than earlier mortality outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Author ContributionsAFH and SKY substantially contributed to the study conception and design. RQZ and JQY conducted the research and prepared the figures. AFH and WJ analyzed and interpreted the data and drafted the article. WWS, SHL, CLY, and LMX critically revised the important intellectual content of the manuscript and provided final approval. AFH and RQZ supervised the study. All authors read and approved the final manuscript.

Funding

This study was sponsored by the Yancheng Science and Technology Bureau (YCBE202365), the Jiangsu Vocational College of Medicine’s School-Local Collaborative Innovation Research Project (202491001), National Key Clinical Specialty, Financial Appropriations of National No. 176 [2022], Flagship Institution of Chinese and Western Medicine Coordination, Financial Appropriations of National No. Jiangsu 60 [2023], Jiangsu Provincial Medical Key Discipline Cultivation Unit (JSDW20221), Yangzhou Medical Key Discipline Cultivation Unit (2024-10), Yangzhou Commission of Health Key Research Project, and the Development Fund Project of Xuzhou Medical University Affiliated Hospital (XYFM202401).

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. The MIMIC-IV database was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA), and consent was obtained for the original data collection.

Consent to participate

All the authors of the article agree to participate in the journal submission.

Consent for publication

All the authors listed have approved the manuscript that is enclose.

Competing interests

The authors declare no competing interests.